Sberbank testing seems to be blown out of proportion by Wikipedia-studying analysts. Or someone else, for some other reason. It's easy to guess that certain “patriotic” circles had decided to make Sberbank support the production by signing some big long term contract, and they decided to NOPE out of it, and quickly made a critical assessment in comparison with top of the line commercial offers. Effective performance of Elbrus in various tests was neither a secret nor a surprise.
The problem isn't performance, it's the architecture. Supposedly, it's intended use case is being a processor for radar control systems with lots of floating point number-crunching/DSP/whatever power to process input data in real-time and also run some general purpose code based on results, all in a single package. Other tasks rarely need that combination of features, it's either top general purpose performance on x86 servers, or top number crunching performance on supercomputers.
Random links:
http://www.elbrus.ru/files/521dd0/256487/1a773c/000000/2014-...
More detailed on page 100 in this book (with many details on other pages, obviously):
http://mcst.ru/doc/book_121130.pdf
Random paper on optimizations:
https://doi.org/10.14529/mmp200109
I remember seeing some benchmark results some time ago in which optimized x86 code running via Elbrus software+hardware translation was faster than native Elbrus compile for the generic C code. Unfortunately, VLIV compilers must be very smart, and ingest all the software and hardware optimizations from others to compete with common architectures.